CNLP study plan
CNLP · Give each week a clear focus.
8 weeks · About 7 hours per week
Natural Language Processing Fundamentals & Text Preprocessing and Feature Engineering
Explain tokenization, stemming/lemmatization, and normalization trade-offs, then implement a tokenizer and a lemmatizer pipeline from scratch on a small corpus
Compute and interpret bag-of-words and TF-IDF representations, including sublinear TF and IDF smoothing variants
Derive and apply n-gram language models with add-k smoothing, then calculate perplexity on held-out text
Build a reusable preprocessing function that handles Unicode normalization, stopword policy, and vocabulary pruning
Your checkpoint
A notebook or repo module that ingests raw text, produces cleaned tokens plus TF-IDF and n-gram features, and reports perplexity before and after smoothing on a held-out split
Classical and Statistical NLP Methods & Natural Language Processing Fundamentals
Implement Naive Bayes and logistic regression text classifiers and compare them on the same feature set
Train a linear-chain CRF or structured perceptron for sequence labeling and report per-entity precision and recall
Explain the bias-variance and generative-versus-discriminative trade-offs as they apply to sparse text features
Interpret confusion matrices and error slices to identify where classical models fail
Your checkpoint
A comparison report with accuracy, macro-F1, and per-class metrics for at least three classical models, plus a short error analysis of 20 misclassified examples
Classical and Statistical NLP Methods & Text Preprocessing and Feature Engineering
Implement word and document embeddings with word2vec-style skip-gram training or a well-understood library, and evaluate nearest neighbors
Apply topic modeling (LDA or NMF) and inspect top terms per topic for coherence
Engineer features from embeddings, POS tags, and entity counts, then measure their marginal contribution
Diagnose and mitigate vocabulary mismatch and out-of-vocabulary handling in a downstream classifier
Your checkpoint
An embedding and topic-model experiment writeup showing nearest-neighbor sanity checks, topic top terms, and an ablation table of engineered feature groups
Machine Learning and Deep Learning for NLP & Large Language Models and Transformers
Build and train an RNN/LSTM text classifier with proper padding, masking, and gradient clipping
Implement scaled dot-product attention and multi-head attention from scratch and verify shapes against a reference
Fine-tune a pretrained transformer encoder for classification and compare against the recurrent baseline
Explain positional encoding, layer normalization, and residual connections in the transformer block
Your checkpoint
A benchmark table comparing recurrent, attention-from-scratch, and fine-tuned transformer models on the same dataset with training curves and hyperparameters documented
Large Language Models and Transformers & Machine Learning and Deep Learning for NLP
Describe encoder-only, decoder-only, and encoder-decoder architectures and match each to appropriate tasks
Explain pretraining objectives, instruction tuning, and parameter-efficient fine-tuning such as LoRA
Design prompts with few-shot examples and measure output stability across paraphrases
Analyze context-window limits, tokenization artifacts, and hallucination risk in a generative pipeline
Your checkpoint
A prompt and fine-tuning experiment log with at least three prompting strategies and one LoRA run, reporting task metrics and a failure taxonomy
Model Evaluation, Ethics, and Responsible AI & Large Language Models and Transformers
Select evaluation metrics appropriate to task type and justify them, including macro versus micro averaging and calibration
Design a human evaluation rubric with annotator agreement measurement
Audit a model for bias across demographic slices and document mitigations
Map a system's data flows and risks to a responsible-AI review, covering privacy, consent, and disclosure
Your checkpoint
An evaluation and ethics dossier containing a metric rationale, an inter-annotator agreement result, a slice-based bias audit, and a risk register with mitigations
Deployment, MLOps, and Production NLP Systems & Model Evaluation, Ethics, and Responsible AI
Package a model behind an API with input validation, timeouts, and structured error responses
Set up experiment tracking, model versioning, and a reproducible training configuration
Configure monitoring for latency, throughput, drift, and data quality with alert thresholds
Write a rollback and shadow-deployment plan for a model update
Your checkpoint
A containerized inference service with a health endpoint, a metrics dashboard definition, and a runbook covering deployment, monitoring, and rollback
Applied NLP Use Cases and Business Integration & Deployment, MLOps, and Production NLP Systems
Translate a business problem into NLP task definitions, success metrics, and acceptance criteria
Estimate cost, latency, and data requirements for the proposed solution and compare build versus buy
Draft a stakeholder-facing summary that states limitations and expected failure modes
Run a full timed practice exam and review every missed item against its source domain
Your checkpoint
A one-page solution proposal for a realistic business use case plus a scored practice-exam review log listing weak domains and remediation actions
Milestones
Week 2: You can implement TF-IDF and an n-gram model from scratch, explain smoothing and perplexity, and justify a classical classifier choice with per-class metrics.
Week 4: You have a working transformer attention implementation and a fine-tuned encoder that beats your recurrent baseline on a documented benchmark.
Week 6: You can defend every metric you report, quantify annotator agreement, and present a slice-based bias audit with concrete mitigations.
Week 8: A deployed, monitored service and a business-facing proposal are complete, and your practice-exam review shows no domain below your target threshold.